Optimizing prediction of key water quality parameters in tilapia river based cage culture system using simple parameters based on different deep learning models

บทคัดย่อ

Importance of the work: Deep learning (DL) models can predict key water quality
values in red tilapia culture using easily measured parameters.
Objectives: To enhance the efficiency and effectiveness of water quality monitoring by
developing a predictive DL model.
Materials and Methods: Convolutional neural network (CNN), long short‒term memory
(LSTM) and a hybrid CNN‒LSTM model were fine‒tuned using the Python software.
Model performance was assessed using root mean square error (RMSE), mean absolute
error (MAE), normalized root mean square error (NRMSE), Nash‒Sutcliffe efficiency
(NSE) and the coefficient of determination (R²). The results were statistically analyzed
based on a t test.
Results: During data collection, the mean ± SD values of water quality parameters were:
dissolved oxygen (DO), 4.03 ± 0.41 mg/L; water temperature (Temp), 27.63 ± 1.42°C;
water pH level (pH), 7.45 ± 0.11; total ammonia nitrogen (TAN) at 0.14 ± 0.04 mg/L;
nitrite‒nitrogen (NO2-‒N), 0.04 ± 0.05 mg/L; alkalinity (ALK), 105.41 ± 9.94 mg/L;
and water transparency (Trans), 75.31 ± 22.80 cm. The study evaluated the CNN, LSTM
and CNN‒LSTM models, with CNN‒LSTM consistently offering the best balance of
accuracy and processing speed. Specifically, it excelled at 1,000 epochs for DO and TAN
predictions and at 2,000 epochs for NO2-‒N and ALK predictions, with no significant
differences compared to observed values using standard measurement methods.
Main finding: The hybrid CNN‒LSTM model that used easily measurable water quality
parameters (Temp, pH and Trans), effectively predicted more difficult‒to‒measure
water quality parameters (DO, TAN, NO2-‒N and ALK). Additionally, the hybrid model
outperformed the individual CNN and LSTM models, providing better prediction
accuracy and faster processing times.